What architecture would you recommend?

Hi Adriana, Good use case, and privacy-driven “keep everything local” requirements come up a lot with ID/document scanning. Both options you listed are viable, so it really comes down to your infrastructure and performance needs. Running in the browser: Roboflow has a JS/web inference SDK that can run models client-side, though depending on how it’s configured it may still call out to Roboflow’s cloud for inference rather than running fully offline in-browser.